Session Summary

The human-readable digest of a session's work. Produced before baton handoff.

Tradeoff: Extra session-end output. Worth it: humans can review and approve
a session's work in 2 minutes instead of re-reading the entire transcript.

---

When to invoke

• At session end (before write-baton.sh runs)
• When the user types /session-summary
• Before any merge to a shared branch

---

Output format

Write to .workflow/state/session-summary.md:

# Session Summary — <session_id>

**Duration**: <start> → <end>
**Tool**: <Claude Code | Codex | Cursor | ...>
**Lifecycle**: <discovery | design | build | harden | launch | operate | evolve>
**Mode**: <greenfield | brownfield | migration | hotfix | refactor | debug-prod | audit>

---

## Tasks completed (verified)

| Task | Spec ref | Claim ID | Confidence | Test exit |
|------|----------|----------|------------|-----------|
| T-014 | specs/auth/tokens.md#ac-3 | claim-2026-05-22-018 | high | 0 |
| T-015 | specs/auth/tokens.md#ac-4 | claim-2026-05-22-019 | high | 0 |
| T-016 | specs/auth/tokens.md#ac-5 | claim-2026-05-22-020 | medium | 0 |

---

## ⚠ Low-confidence items (require human review)

- claim-2026-05-22-020 (T-016): expiry test passes but uses fake timers.
  Real-clock verification recommended in staging.

---

## 🚧 Open blockers

(none)

OR

1. **AC-7 ambiguity** — should refresh tokens be invalidated on password change?
   Need spec clarification. See `.workflow/state/blockers.md`.

---

## Files modified

- `src/auth/refresh.ts` (added expiry handling)
- `src/auth/refresh.test.ts` (added 3 tests)
- `specs/auth/tokens.md` (updated AC-3 wording with user approval)

---

## Cost

- LLM spend: $X.XX
- Per-model breakdown: Haiku $X.XX, Sonnet $X.XX, Opus $X.XX
- Per-subagent breakdown: planner $X.XX, coder $X.XX, ...

---

## Next session

The baton at `.workflow/state/baton.md` contains the literal `next_action`.
The next agent should read it first.

---

## Constitution adherence

- Spec-anchoring: ✓ all changes traced
- Verified claims: ✓ 3/3 verified
- Surgical edits: ✓ no out-of-scope changes
- Budget: ✓ within session cap
- Loop detection: ✓ no loops
- Client boundary: ✓ no cross-project access

---

Verification before signing

Before marking the session "complete":

1. Run claim-validator on all proofs. Confirm 100% pass.
2. Check blockers file. If any are unresolved AND the session changed code that
   touches them, flag it.
3. Recompute the cost from budget-ledger.jsonl.
4. Verify spec-trace-rate is 100%.

If any check fails, the summary is marked not safe to merge and the agent
must escalate to the user.

---

This skill is working when: humans approve sessions in < 2 minutes based on
the summary, without re-reading the transcript.